[PATCH] D115713: [LV] Don't apply "TinyTripCountVectorThreshold" for loops with compile time known TC.

[Dave Green via Phabricator via llvm-commits]

dmgreen added a comment.

> Avoid vectorization if there is reduction/recurrence induced overhead

For some architectures like MVE and AArch64 the reduction cost might well be cheap. Non-zero but also not huge. It can depend on the reduction. Have you considered the cost of constants and splats too, that the vectorizer will currently hoist into the preheader? The costmodel functions might also sometimes return cheaper costs for certain shuffles and gathers under the assumption that loop invariant parts can be hoisted.

I've tried to get an example of the case that is getting worse for MVE in rG662b5f18467e <https://reviews.llvm.org/rG662b5f18467e7e2dee2d07358efecee349c4b2e7>. It might be a bit over-reduced, but hopefully it still shows the same problems. The original I can't share, I'm afraid.

@fhahn did you have an example of cases where you had seen AArch64 getting worse? Is it covered by reductions or is it more general?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D115713/new/

https://reviews.llvm.org/D115713